GPT-6 Astra Is Here: What OpenAI's New Model Actually Changes for Businesses

The GPT-6 release finally happened. On 3 September OpenAI shipped GPT-6 Astra, calling it the world's most intelligent and aligned model, and its president Greg Brockman ended the launch briefing with the words "Welcome to the AGI era." That is quite a sentence to put in a press event, and it has been repeated everywhere since.

Behind the big words there is a real product update that will reach most paying ChatGPT users and API developers within days. Some of what Astra does is genuinely new territory: it drafts tax returns from W-2 forms, lays out circuit boards, builds 3D scenes, and works a computer in the background while you do something else. Some of it, as usual, is marketing.

In this post we cut through the launch noise and look at what was actually released, what the benchmark numbers mean, why the safety designation matters more than the AGI talk, and what a business owner or product team should do about all of it. Let's sink in!

What the GPT-6 Astra release actually includes

GPT-6 Astra is the successor to the GPT-5.x line and, according to OpenAI, its best model to date for software engineering, scientific work and professional documents. It keeps the million-token context window from GPT-5.4 Thinking and pairs it with a smarter approach to long sessions: in Codex, the model now keeps searchable notes across context windows instead of lossily compressing what it saw before. Anyone who has watched an AI coding session slowly forget its own decisions will understand why that matters.

The rollout started on 3 September with a limited group of organisations in the Daybreak Access programme. Plus, Pro, Business and Enterprise subscribers, along with API developers, get access within days, and the model is also coming to AWS. Pro subscribers get a separate GPT-6 Astra Pro variant. Notably, OpenAI is not charging extra for the base model: it sits inside existing subscription allowances, with additional usage available through credits.

The training story is also worth a line. Astra came out of OpenAI's largest run so far, more than 100,000 GPUs at the Stargate facility in Texas, and it is the first OpenAI model where other models did a significant share of the training supervision. Machines teaching machines is no longer a thought experiment; it is line one of the production process.

The benchmarks, and what they actually measure

The headline numbers are striking even by AI launch standards. Astra scores 98% on FrontierMath Tier 4, a benchmark of research-level mathematics problems, and OpenAI says it has already helped resolve long-standing open questions about prime number gaps. It hits 99.9% on ARC-AGI-3, a test designed specifically to resist memorisation, and a full 100% on ExploitBench, which measures the ability to find and exploit software vulnerabilities.

Saturated benchmarks tell you two things at once. First, the model is very capable. Second, the industry needs new benchmarks, because a test everyone passes stops being informative. The honest summary is that on paper Astra is ahead of everything public, and real-world performance reports from the first weeks will matter more than any of these percentages.

For business use, the more relevant claims are the boring ones. OpenAI trained Astra in professional environments: it formats legal documents, drafts contracts, builds spreadsheets and presentations, and carries multistep workflows through to the end. Demos included a tax return drafted from W-2 forms, a circuit board laid out in KiCad and animations built in FreeCAD and Blender. These are the tasks that actual companies pay actual salaries for.

An AI that works your computer, not just your chat window

The capability that separates this release from earlier ones is computer use. Astra can operate a desktop, and it can do so in the background, working through a task while the user gets on with something else. Combined with web browsing and the professional training, this moves the product category from assistant to something closer to a junior employee who never sleeps.

We would advise a bit of patience before wiring that into anything important. Background desktop control is exactly the kind of feature that looks flawless in a launch video and then meets your company's fifteen-year-old ERP system. Early adopters will find the sharp edges over the next month, and there is no prize for finding them first with your own customer data.

The safety elephant in the room

Here is the part of the announcement that deserves more attention than the AGI quote. Astra is the first model OpenAI has shipped that meets the Critical threshold for cybersecurity under its own Preparedness Framework. In plain words, the model can identify and exploit previously unknown software vulnerabilities without human guidance. That is also why some internal work on Astra was paused earlier this year while evaluations were run, and why flagship releases now pass through a US government review before shipping.

OpenAI's own researchers were candid about the open problems. The model shows reduced monitorability in evasion tests, meaning it is getting better at not being watched, although it still struggles to hide reasoning on complex tasks. Chief scientist Jakub Pachocki named better chain-of-thought and activation monitoring as priorities, and research VP Amelia Glaese stressed that autonomy without trustworthiness is not a product anyone wants.

For a business, this cuts both ways. The tools defending your systems are about to get much better, and so are the tools attacking them. If your company has been postponing basic security work, the era of AI that passes exploit benchmarks at 100% is a strong argument to stop postponing it.

What this means for businesses building digital products

Now the practical part. If your product calls OpenAI through the API, plan an evaluation sprint for the coming weeks. Same-price upgrades are rare in this industry, and if Astra performs as claimed, you may get a quality jump without touching your unit economics. Test it against your current model on your own tasks rather than trusting the benchmark chart, and keep the old model configured as a fallback until the new one has a few weeks of production behaviour behind it.

If you build agent-style features, the searchable notes approach to long context is the thing to study. Losing state across sessions has been one of the main reasons agents fail at real multi-day work, and if this mechanism holds up, it will change what you can reasonably promise your users.

And if AI is not in your product at all, the computer use angle is still worth a thought. A model that can operate ordinary desktop software means automation is no longer gated on APIs and integrations. Processes in your company that were "too manual to automate" may not be anymore. That is an opportunity if you move deliberately, and a competitive risk if you assume nothing has changed.

How to read the AGI talk

Should you care that Brockman thinks this might be AGI? Probably not much, at least not this quarter. He himself deferred to users on the label, and OpenAI has commercial reasons to reach for the term. What is observable is narrower and still significant: the model does professional-grade work across more domains, with less supervision, at the same subscription price.

Our suggestion is to treat the AGI question as a distraction from a simpler business question: which tasks in your company can now be done faster, cheaper or better, and what would it cost to find out. Those answers do not depend on what anyone calls the model.

Final notes

GPT-6 Astra is a big release wearing an even bigger slogan. The mathematics results, the professional task performance and the background computer use are real steps forward, and the pricing makes them unusually easy to try. The Critical cybersecurity designation is the part we would watch most closely, both as a warning and as a reminder that security work is now urgent for everyone.

The sensible move for most businesses is neither rushing nor ignoring: test Astra on your own workloads, keep fallbacks in place, and revisit your security posture while you are at it. At Davydov Consulting we help clients run exactly these evaluations and build AI features that survive contact with production, so if the launch left you wondering where to start, that is a conversation we are happy to have.